🎯 Top 3 Things to Know
1. OpenAI launched a $14 billion "Deployment Company" backed by $4 billion in initial capital from 19 firms, and made Edinburgh-based Tomoro its founding acquisition. The pitch is unsubtle. Capability is no longer the bottleneck. Adoption is. The subsidiary copies Palantir's Forward Deployed Engineer playbook: 150 engineers, mostly from Tomoro, will sit inside customer organizations and rebuild workflows around the API rather than ship slides. TPG leads the partnership. Bain, Capgemini, and McKinsey are on the cap table alongside Goldman Sachs and SoftBank. Frontier labs are now budgeting for services revenue, not just inference revenue, which puts them in direct competition with the systems integrators they just funded. Worth watching whether Anthropic responds with its own services arm. OpenAI announcement · Bloomberg
2. Google reframed Android as an "intelligence system" at the Android Show I/O Edition, with Gemini Intelligence acting across apps and a new Googlebook laptop line shipping this fall. The interesting move is not the chat window. It is that Gemini now reads what is on screen, decides which app to invoke, and completes the task without the user switching surfaces. Sameer Samat said Google is "transitioning from an operating system to an intelligence system." Cross-app agent behavior rolls out first on Pixel and Samsung Galaxy this summer, then watches, cars, glasses, and the new Googlebook line built with Acer, Asus, Dell, HP, and Lenovo. Apple's WWDC arrives June 8 with its own Siri overhaul. The contested surface for the next year is the OS-level agent, not the model. Google blog · CNBC
3. Anthropic released Claude for Small Business, a packaged bundle of connectors and prebuilt workflows targeting Quickbooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365. SMB has been the missing segment in frontier-lab go-to-market. Enterprise gets bespoke deployment, consumers get the chat app, and the long tail of small operators has been buying Claude indirectly through whichever vertical SaaS reached them first. The Small Business bundle is the first attempt to ship Claude as a configured workflow product rather than an API surface. The competitive read is that Microsoft Copilot and Google Workspace have a multi-year head start with this audience through bundled distribution, so Anthropic's bet is that prebuilt cross-tool agents outweigh native integration. Worth watching adoption among accounting and bookkeeping firms, where the Quickbooks plus Docusign plus Gmail surface is the daily workflow. Claude for Small Business
🚀 Frontier Models & Features
Anthropic also added new tooling to Claude Managed Agents this week, including a research-preview "dreaming" feature that lets agents review prior sessions for patterns, plus a multiagent orchestration tool that lets a lead agent delegate subtasks to specialists working in parallel on a shared filesystem. The orchestration primitive matters because it is the first first-party answer to the harness layer that production teams have been writing themselves. Anthropic news
🔬 Research Worth Reading
LaTER: Efficient Test-Time Reasoning via Latent Exploration and Explicit Verification (Li, Wang, Liu et al.). arXiv 2605.07315
- TL;DR: Two-stage reasoning where the model first explores candidate paths in a continuous latent space, then switches to explicit chain-of-thought only for verification and final answer generation. The expensive token-by-token reasoning happens only after the search has narrowed.
- Stat: On Qwen3-14B, training-free LaTER cuts total token usage by 16 to 32 percent across several benchmarks while matching or improving accuracy. AIME 2025 goes from 70.0 to 73.3 percent on 10,661 tokens instead of 15,730.
- Apply it: If your reasoning workload is dominated by output token cost, the latent-then-verify pattern is worth a measured pilot. Run it side by side with full CoT on a sample of your hardest queries and compare cost per resolved query, not just accuracy.
Beyond the Black Box: Interpretability of Agentic AI Tool Use (Tatsat & Shater). arXiv 2605.06890
- TL;DR: A practical interpretability toolkit using sparse autoencoders and linear probes that reads model states before each tool call to predict both whether a tool is needed and how consequential the next action is.
- Stat: The probes flag high-consequence tool calls before they execute, providing an internal signal that decouples from the model's natural-language explanation of its own actions.
- Apply it: If your agent runtime treats every tool call the same, this is the cleanest published recipe for a pre-execution risk signal that does not depend on the model self-reporting. Pair it with a policy layer like AgentTrust and you have a two-stage gate: predict consequence, then decide whether to allow.
🏢 Enterprise in the Wild
OpenAI Chief Revenue Officer Denise Dresser said enterprise revenue is at a "tipping point," with enterprise now above 40 percent of total revenue and expected to reach parity with consumer by year end. The data point worth the headline is not the percentage. It is the rate of change. A year ago consumer dominated by roughly two to one. Forty percent in twelve months is a steeper enterprise ramp than Salesforce, Workday, or ServiceNow posted in their fastest comparable years. CNBC
🛠️ Tooling & Ecosystem
The MCP Registry crossed 2,000 published server entries this month, roughly eight months after launch in September 2025. SDKs in Python, TypeScript, C#, and Java now report a combined 97 million monthly downloads. The April Dev Summit in New York drew 1,200 attendees, and the protocol has been operating under the Linux Foundation's Agentic AI Foundation since December. The interoperability story is now boring in the best sense. WorkOS overview
⚖️ Policy & Regulation
The EU reached political agreement on May 7 on the AI Act omnibus amendments, with transparency rules effective August 2026 and high-risk system rules pushed to December 2, 2027. The slip is consequential. Biometrics, employment screening, education, critical infrastructure, and border control systems get an additional year before binding obligations. The EU framing is "implementation realism." The competitive read is that EU and U.S. timelines now diverge further: the U.S. December 2025 executive order is actively preempting state laws while the EU softens its own. Artificial Intelligence Act tracker
📌 Watch List
- Frontier labs becoming systems integrators, and the channel conflict it creates with Bain, Capgemini, and McKinsey.
- OS-level agents (Gemini Intelligence, the rumored Siri overhaul) replacing standalone assistant apps as the consumer AI surface.
- SMB as the next contested segment, packaged-workflow bundles versus Microsoft and Google's distribution moat.
- Latent-space reasoning as a cost lever for production CoT workloads.
- EU AI Act timeline slippage and what it implies for enterprise compliance planning in regulated sectors.